Opium poppy cultivation mapping by phenological signature
Opium poppy produces a spectral trajectory through flowering and capsule stages that no cereal crop replicates. Time-series analysis of red-edge and EVI signals on Sentinel-2 and PlanetScope can map cultivation at field scale, though cloud cover and plot fragmentation set real limits on what satellites can count.
Sensors
- Sentinel-2 MSI (ESA): 10 m resolution in visible and near-infrared bands; 20 m in red-edge bands (B5, B6, B7) that are central to NDRE computation. Five-day revisit at the equator with both satellites active. Free archive back to 2015. Cloud contamination during the South Asian monsoon (June–September) can eliminate entire growing-season windows in Helmand and Shan State.
- PlanetScope SuperDove: 3 m resolution, daily revisit globally, eight spectral bands including a dedicated red-edge channel at 705 nm. Commercial licence required. Sub-field detail resolves plots down to roughly 0.1 ha, which matters for the fragmented smallholder terraces of Myanmar's Shan State. Archive depth varies by area of interest and licence tier.
- Landsat 8/9 OLI (USGS/NASA): 30 m resolution, 16-day revisit per satellite (8-day combined). No dedicated red-edge band, so NDRE is not directly computable; EVI and NDVI time-series are the primary indices. Free archive back to 1972 for Landsat legacy missions, 2013 for OLI. Useful for multi-year trend analysis and baseline establishment.
- MODIS Terra/Aqua (NASA): 250 m to 500 m resolution, daily revisit. Too coarse to map individual plots but useful for regional phenological calendars and for flagging anomalous green-up events at province scale that cue tasking of higher-resolution sensors. MODIS EVI time-series underpins several published UNODC-adjacent crop-calendar studies.
What the flowering window gives away
Opium poppy (Papaver somniferum) flowers for roughly ten to twenty days, depending on variety and altitude. During that window its canopy reflectance in the red-edge region (approximately 700–740 nm) spikes sharply above background soil and above the wheat fields that typically surround it. The Normalised Difference Red Edge index (NDRE = (NIR – RedEdge) / (NIR + RedEdge)) captures this spike clearly on sensors with a dedicated red-edge band. On Sentinel-2, Band 5 at 705 nm and Band 8A at 865 nm are the standard pair. The Enhanced Vegetation Index (EVI) adds a soil-adjustment term that reduces noise in the thin-canopy early stages, making it useful for detecting emergence before flowering.
The phenological calendar is the key discriminator. In Afghanistan's Helmand basin, planting runs from October to December and flowering peaks in March and April. In Myanmar's Shan State the calendar shifts roughly six weeks later. Mexico's Sierra Madre Occidental sees multiple staggered harvests across altitudinal zones. A classifier trained on a single calendar will miss out-of-season plots. Published UNODC survey methodology documents these regional calendars explicitly, and any serious mapping programme must parameterise its detection window accordingly.
Separating poppy from wheat: the time-series argument
A single-date spectral image is rarely sufficient. Wheat and poppy share similar NDVI values during early vegetative growth. The separation emerges in the trajectory: poppy's red-edge reflectance peaks earlier than wheat's heading stage and declines faster as the petals drop and the capsule matures. The capsule stage itself produces a distinctive dull-green to grey-green signal that is spectrally flatter than ripening wheat. Capturing this full arc requires at least four to six cloud-free acquisitions between emergence and harvest, which is achievable in Afghanistan's dry spring but genuinely difficult in Myanmar during the tail of the northeast monsoon.
Random forest and support vector machine classifiers applied to multi-temporal NDRE and EVI stacks are the published standard. A 2019 study in Remote Sensing (MDPI) demonstrated greater than 85 percent overall accuracy for poppy mapping in Helmand using Sentinel-2 time-series, with the main confusion class being fallow fields that briefly greened after rain. Object-based image analysis, which groups spectrally similar pixels into field-shaped objects before classification, reduces salt-and-pepper noise and is particularly important at 10 m resolution where a single poppy plot may span only a few dozen pixels.
Where the method fails, and by how much
Cloud cover is the primary operational constraint. The UNODC Afghanistan Opium Survey has historically noted that cloud contamination can render 10 to 30 percent of the survey zone uninterpretable in any given year. Synthetic aperture radar (SAR) from Sentinel-1 can penetrate cloud but its backscatter signal does not carry the spectral phenological information that makes poppy separable from wheat. SAR contributes crop-height and canopy-structure proxies but cannot substitute for optical time-series in the classification step.
Resolution sets the minimum detectable plot size. At Sentinel-2's 10 m, a plot smaller than roughly 0.1 ha (a 30 m by 30 m patch) will be mixed with surrounding soil or adjacent crops in most pixels, causing systematic undercounting of fragmented smallholder cultivation. PlanetScope at 3 m pushes this floor down to plots of perhaps 0.01 ha, but commercial tasking costs and data-volume constraints limit how much area can be covered at that resolution. UNODC survey reports consistently note that smallholder plots in Shan State are underrepresented relative to ground-truthed estimates, and this is the principal reason.
Deliberate countermeasures are a real but poorly documented factor. Growers in some regions intercrop poppy with wheat or plant under orchard canopies specifically to reduce spectral separability. These tactics are not hypothetical. They shift the problem from remote sensing into field verification, which satellites cannot replace.
Cross-referencing UNODC survey methodology
The UNODC annual opium surveys for Afghanistan and Myanmar use a stratified random sample of ground-truth points, aerial photography and satellite imagery in combination. Their published methodology reports are the most rigorous public reference for what satellite detection can and cannot confirm at national scale. The surveys report cultivation area in hectares with confidence intervals, not point estimates, precisely because of the cloud and resolution limits described above.
An independent mapping programme should treat UNODC estimates as a calibration benchmark rather than a ground truth to be reproduced. Where the satellite estimate diverges from the UNODC figure, the divergence is informative: it may indicate new cultivation in areas not sampled by the survey, or it may indicate classification error. Distinguishing between those two explanations requires field verification or very-high-resolution imagery, neither of which satellites can provide automatically.
Building an operational detection pipeline
A practical pipeline has four stages. First, define the detection window from the regional planting calendar and pull all available Sentinel-2 scenes for the target province across that window, flagging cloud-masked pixels. Second, compute NDRE and EVI for each scene and build per-pixel time-series stacks, interpolating sparingly across short cloud gaps (one to two scenes) using harmonic fitting. Third, apply a trained classifier, ideally one validated against the most recent UNODC ground-truth sample for that region. Fourth, produce a change layer against the prior-year baseline to flag new cultivation areas for priority follow-up.
Satellize applies this pipeline architecture to open Sentinel and Landsat archives and can add commercial PlanetScope tasking under client licence for sub-field resolution in priority zones. The crop-estimation work done for the Kingdom of Tonga demonstrates the phenological time-series approach in a smallholder context, though the crop and the policy question differ substantially. For enforcement clients, the deliverable is a georeferenced cultivation map with per-plot area estimates, confidence scores and a change-detection layer, updated at each growing season.
Typical figures
| Spatial resolution (primary) | 10 m (Sentinel-2 visible/NIR); 20 m (Sentinel-2 red-edge bands); 3 m (PlanetScope SuperDove); 30 m (Landsat OLI) |
| Revisit frequency | 5 days (Sentinel-2 combined); daily (PlanetScope); 8 days combined (Landsat 8+9); daily (MODIS) |
| Key spectral bands | Red-edge ~705 nm (Sentinel-2 B5, PlanetScope Band 6); NIR ~865 nm (Sentinel-2 B8A); Red ~665 nm (B4); Blue ~490 nm (B2) for EVI |
| Primary indices | NDRE = (NIR – RedEdge) / (NIR + RedEdge); EVI = 2.5 × (NIR – Red) / (NIR + 6×Red – 7.5×Blue + 1) |
| Minimum detectable plot size | ~0.1 ha at 10 m (Sentinel-2); ~0.01 ha at 3 m (PlanetScope); fragmented plots below these thresholds are systematically undercounted |
| Detection window (Afghanistan) | March–April (flowering peak); 4–6 cloud-free acquisitions required across the window for reliable classification |
| Cloud contamination risk | Low in Afghan spring; moderate to severe in Myanmar Shan State (November–February); up to 30% of survey zone may be uninterpretable in a poor year |
| Archive depth | Sentinel-2 from 2015; Landsat OLI from 2013; Landsat legacy from 1972; PlanetScope varies by licence |
| Delivery formats | GeoTIFF cultivation mask; GeoPackage per-plot polygons with area and confidence attributes; CSV change-detection summary; PDF seasonal report |
| Classification accuracy (published benchmark) | >85% overall accuracy reported for Sentinel-2 time-series in Helmand (published Remote Sensing literature); lower in fragmented smallholder landscapes |
Analytics Satellize can run
| Seasonal cultivation area estimate | Multi-temporal NDRE/EVI stack classified by random forest or SVM, parameterised to regional planting calendar | GeoTIFF cultivation mask and per-province hectare totals with confidence intervals, delivered at end of flowering window |
| Year-on-year change detection | Pixel-wise comparison of classified cultivation masks between consecutive seasons; new and abandoned plot polygons flagged | GeoPackage change layer with new-cultivation and abandonment polygons, ranked by area for priority follow-up |
| Phenological anomaly alert | MODIS EVI time-series monitored at province scale for early-season green-up anomalies that trigger high-resolution tasking | Automated alert with bounding box and recommended Sentinel-2 or PlanetScope tasking window, delivered within 48 hours of anomaly detection |
| Sub-field plot delineation | Object-based image analysis on PlanetScope 3 m imagery; field boundaries extracted by spectral and shape segmentation | Per-plot polygon layer with estimated area, NDRE peak value and phenological stage at time of acquisition |
| Cloud-gap assessment and data-quality report | Scene-by-scene cloud-mask analysis across the detection window; percentage of target area with sufficient cloud-free observations quantified | Data-quality report flagging districts with insufficient temporal coverage and recommending supplementary commercial tasking |
| Multi-year trend analysis | Landsat OLI EVI/NDVI time-series from 2013 to present; cultivation area estimated annually for trend and displacement analysis | Time-series chart and tabular dataset by district, suitable for policy reporting or UNODC cross-reference |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.